Bot gives wrong answers

The most common cause of off-target answers is a gap or inconsistency in the knowledge base, not the language model itself. The diagnostic path below will help you quickly identify the source of the problem.

1. Identify gaps in the knowledge base

Open the Control center. In the Escalation reasons table you'll see which topics are most often not covered in the knowledge base. If a category name matches the wrong answer, adding even a single article on that topic usually solves the problem.

In the same tab, AI insights automatically suggest topics to cover based on recent conversations.

2. Check the bot's confidence

Low confidence means the model "wasn't sure" of the answer and probably improvised. In the Control center, on the Confidence distribution chart, you'll see what share of conversations have low confidence.

If the share of low confidence is high:

• the knowledge base is too shallow, so add specific articles,
• the articles are too general, so add facts (see Best practices),
• the articles mix topics together, so split them into smaller ones.

3. Fix or remove problematic articles

If you suspect the bot is relying on a specific article that's misleading:

• open it in the Documentation tab and fix the content so it's precise and matches the current state of your offering,
• or delete the article if the topic no longer applies to you or the content is so outdated that it's easier to write a new entry.

After editing or deleting, the change takes effect in the next conversation with the bot.

4. Adjust the knowledge base context

In Bot configuration, try reducing the Context depth to 1 if the bot is pulling in too many unrelated articles. Or increase it to 3 if you notice that the answers are too shallow.

5. Check the persona and brand description

Check Brand and Messages:

• whether the brand description is misleading (e.g. it covers products you don't actually sell),
• whether the bot's personality causes it to resort to jokes or improvisation in situations that call for being factual.

When to write to us

If you've gone through all the points above and the bot still gets specific facts wrong, write to us at kontakt@fibly.pl and include:

• an example of a conversation with the wrong answer (a link to the dashboard),
• the content of the knowledge base article that should be the source,
• what specifically you expected from the bot.

We'll help diagnose whether the problem lies in the knowledge base, the model, or the configuration itself.